While importing the `ouroboros` crate and the `aliasable` crate it
depends on, the "unsafe Rust reviewer" expressed some concern that
they contain a lot of unsafe code that's hard to review. We can avoid
the unsafe code altogether by making `TreeEntriesIterator` not
self-refential. Instead, we can collect the matching entries in an
individual tree up front. It does have some performance cost:
```
❯ hyperfine --warmup 3 --runs 30 \
'/tmp/jj-before --ignore-working-copy files -r v6.0' \
'/tmp/jj-after --ignore-working-copy files -r v6.0'
Benchmark 1: /tmp/jj-before --ignore-working-copy files -r v6.0
Time (mean ± σ): 461.4 ms ± 14.3 ms [User: 232.1 ms, System: 229.4 ms]
Range (min … max): 443.4 ms … 496.3 ms 30 runs
Benchmark 2: /tmp/jj-after --ignore-working-copy files -r v6.0
Time (mean ± σ): 482.0 ms ± 14.3 ms [User: 257.2 ms, System: 224.9 ms]
Range (min … max): 461.8 ms … 513.3 ms 30 runs
Summary
'/tmp/jj-before --ignore-working-copy files -r v6.0' ran
1.04 ± 0.04 times faster than '/tmp/jj-after --ignore-working-copy files -r v6.0'
```
I think that's acceptable.
This is much faster (maybe because of better cache locality?) Another option
is to use BTreeSet, but the BinaryHeap version is slightly faster.
"bench revset" result in my linux repo:
revsets/heads(tags())
---------------------
baseline 3.28 560.6±4.01ms
1 2.92 500.0±2.99ms
2 1.98 339.6±1.64ms
3 (this) 1.00 171.2±0.30ms
Apparently, IndexEntry::generation_number() isn't cheap probably because it
involves random access to larger memory region, and the u32 value might not
be aligned. Let's instead store the generation numbers in BinaryHeap.
Also, heads_pos() becomes slightly faster by keeping the BinaryHeap entries
small, so I've removed the IndexEntry at all.
This makes the default log and disambiguation revsets fast, which evaluate
'heads(immutable_heads())'.
"bench revset" result in my linux repo:
revsets/heads(tags())
---------------------
baseline 3.28 560.6±4.01ms
1 2.92 500.0±2.99ms
2 (this) 1.98 339.6±1.64ms
All callers just iterate over the parent entries.
"bench revset" result in my linux repo:
revsets/heads(tags())
---------------------
baseline 3.28 560.6±4.01ms
1 (this) 2.92 500.0±2.99ms
For loose refs, uninteresting directories can be just skipped. For packed refs,
gix will have to do binary search for each prefix to find the starting point.
Still it's better overall if the repository contains tons of refs/jj/keep refs.
With my linux repo containing ~5k loose jj refs, this saves ~40ms:
% hyperfine --warmup 3 --runs 10 \
"/tmp/jj-gix --ignore-working-copy git import -R ~/mirrors/linux" \
"/tmp/jj-gix-iter --ignore-working-copy git import -R ~/mirrors/linux"
Benchmark 1: /tmp/jj-gix --ignore-working-copy git import -R ~/mirrors/linux
Time (mean ± σ): 151.6 ms ± 11.4 ms [User: 38.8 ms, System: 111.6 ms]
Range (min … max): 129.8 ms … 159.5 ms 10 runs
Benchmark 2: /tmp/jj-gix-iter --ignore-working-copy git import -R ~/mirrors/linux
Time (mean ± σ): 109.9 ms ± 11.6 ms [User: 27.5 ms, System: 82.4 ms]
Range (min … max): 89.4 ms … 117.8 ms 10 runs
Gitoxide errors are boxed since there are various error types and they tend
to exceed the clippy size limit.
Apparently, gitoxide is faster than git2:
% hyperfine --warmup 3 --runs 10 \
"/tmp/jj-baseline --ignore-working-copy git import -R ~/mirrors/linux" \
"/tmp/jj-gix --ignore-working-copy git import -R ~/mirrors/linux"
Benchmark 1: /tmp/jj-baseline --ignore-working-copy git import -R ~/mirrors/linux
Time (mean ± σ): 205.4 ms ± 15.7 ms [User: 59.6 ms, System: 144.6 ms]
Range (min … max): 189.7 ms … 223.9 ms 10 runs
Benchmark 2: /tmp/jj-gix --ignore-working-copy git import -R ~/mirrors/linux
Time (mean ± σ): 176.2 ms ± 13.7 ms [User: 41.2 ms, System: 134.0 ms]
Range (min … max): 155.4 ms … 186.5 ms 10 runs
If a commit pointed to by HEAD or ref is missing, the ref is considered
invalid and excluded by import_refs(). The current test behavior appears to
depend on some in-memory cache of git2::Repository.
We need to .collect_vec() the parents iterator to temporary buffer since the
borrowed iterator can't be returned back to the dag_walk functions. Another
option is to clone op_store and parent ids to remove &self lifetime from the
iterator, but that also means a temporary Vec is created.
Unlike dfs_ok(), this function short-circuits at an Err as we use non-lazy
topo_order_forward() internally. I think that's good enough. If we implement
GC on operation log, deleted parents will be excluded (or mapped to tombstone)
by caller. An Err shouldn't mean it's GC-ed.
This unblocks the use of Result<T, E> in op.parents().
There are two ways to encode errors:
a. impl IntoIterator<Item = Result<T, E>>
b. Result<V, E> where V: FromIterator<Item = T>
I think (a) is more natural to algorithms like dfs(), which can process error
nodes transparently.
Still the caller might have to collect the source iterator to temporary Vec
to conform to the neighbors_fn signature. It's not easy for neighbors_fn to
return an iterator borrowing the input node. We already have GAT, but doesn't
have return-position impl Trait in trait yet.
Recognize signature metadata from git commit objects, implement
a basic version of that for the native backend.
Extract the signed data (a commit binary repr without the signature) to
be verified later.
Otherwise, ref updates would fail if we port git::export_refs() to gitoxide.
This change isn't strictly needed for the backend itself, but we'll reuse the
gix::Repository instance created by the backend when importing and exporting
Git refs.
GitBackend will use it to configure gix::Repository. I think UserSettings
is generally useful to pass store-specific parameters, so I've updated all
factory functions.
While the safe implementation is a bit more complex (and probably more branchy),
I don't think the runtime overhead would matter here. Let's remove one more
unsafe for better code maintainability.
We have a few places where we have a `MergedTreeValue` and need to
read the data associated with it so we can write to the working copy
or include it in a diff. Let's extract some of that shared logic to a
function so we can reuse it. I plan to use it for reading file
contents in advance while streaming a diff in `local_working_copy`
soon (and probably in `jj diff` thereafter), but I think it seems like
an improvement on its own.
I'd like to read N files ahead from the backend, to avoid serializing
too many server calls on backends that are backed by a server. Moving
the reads a little earlier is a little step towards that.
The `TreeState::write_*()` functions can now be made into free/static
functions if we prefer.
We no longer have "unsafe" in this function, so let's use the iterator API
instead of recursion. Apparently I haven't pushed this change before because
unsafe in .find_map() looked scary.
Remove a couple of unnecessary unsafes:
- The NonZeroUsize is a constant where the unwrap will optimize away
anyway and we don't have an unsafe without any good reason there :)
- The other two were simply not needed, lifetimes worked fine, maybe
Rust became better since that code was written? NLL? Anyway, they're
gone now
As discussed in Discord, it's less useful if remote_branches() included
Git-tracking branches. Users wouldn't consider the backing Git repo as
a remote.
We could allow explicit 'remote_branches(remote=exact:"git")' query by changing
the default remote pattern to something like 'remote=~exact:"git"'. I don't
know which will be better overall, but we don't have support for negative
patterns anyway.
It's super common that a Merge object holds a resolved value, so let's inline
up to 1 element. T of Merge<T> usually consists of a couple of pointer-sized
fields. I don't see any measurable speed up, but it's no worse than the
original.
Since the concurrent diff algorithm is significantly slower when using
the Git backend, I think we'll have to use switch between the two
algorithms depending on backend. Even if the concurrent version always
performed as well as the sequential version, exactly how concurrent it
should be probably still depends on the backend. This commit therefore
adds a function to the `Backend` trait, so each backend can say how
much concurrency they deal well with. I then use that number for
choosing between the sequential and concurrent versions in
`MergedTree::diff_stream()`, and also to decide the number of
concurrent reads to do in the concurrent version.
When diffing two trees, we currently start at the root and diff those
trees. Then we diff each subtree, one at a time, recursively. When
using a commit backend that uses remote storage, like our backend at
Google does, diffing the subtrees one at a time gets very slow. We
should be able to diff subtrees concurrently. That way, the number of
roundtrips to a server becomes determined by the depth of the deepest
difference instead of by the number of differing trees (times 2,
even). This patch implements such an algorithm behind a `Stream`
interface. It's not hooked in to `MergedTree::diff_stream()` yet; that
will happen in the next commit.
I timed the new implementation by updating `jj diff -s` to use the new
diff stream and then ran it on the Linux repo with `jj diff
--ignore-working-copy -s --from v5.0 --to v6.0`. That slowed down by
~20%, from ~750 ms to ~900 ms. Maybe we can get some of that
performance back but I think it'll be hard to match
`MergedTree::diff()`. We can decide later if we're okay with the
difference (after hopefully reducing the gap a bit) or if we want to
keep both implementations.
I also timed the new implementation on our cloud-based repo at
Google. As expected, it made some diffs much faster (I'm not sure if
I'm allowed to share figures).
I'm about to add a few more checks for diffing with a matcher. I think
it will help make it readable and reduce the risk of mixing up
variables between each part of the test if we use some nested blocks.
I also removed some unnecessary `.clone()` calls while at it.
